Thermostats refactoring

[jngrad]

Description of changes:

• encapsulate thermostats and remove their Cython interface
  • use one C++ ScriptInterface class per core class (fixes Split thermostats into separate classes #3980)
  • introduce the ThermostatFlags enum, allow deactivating a single thermostat, add an is_active flag accessible from the Python interface (fixes Refactor thermostat switches #4266)
• remove 20 MPI callbacks
  • partial fix for Get rid of mpi callbacks #4614
• remove thermostat and thermalized/rigid bond global variables
  • partial fix for Remove global variables #2628
• remove a significant amount of historical Cython code that relied on deprecated Cython features
  • partial fix for De-cythonize ESPResSo #4542
  • ESPResSo is now compatible with Cython 3.0
  • the nogil deprecation warning subsists
• improve testing of the thermostats interface
• API changes:
  • thermalized bond parameter seed was moved to system.thermostat.set_thermalized_bond()
    • this change better reflects the fact there is only one global seed for all thermalized bonds
    • until now this global seed was overwritten by any newly created thermalized bond, whether it was added to the system or not
  • LB thermostat seed parameter now gets written to the RNG seed (silent change)
    • until now, the RNG seed was hardcoded as 0 and the seed value was used to set the RNG counter, thus two independent simulations with a different seed would actually get the same particle coupling random noise sequence with a time lag equal to the difference between the two seeds
    • this is a bugfix for ensemble runs
    • fixes LB particle coupling thermostat seed doesn't fully decorrelate independent simulations #4848

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[jngrad]

Regarding the API change for thermalized bonds: the thermalized bond thermostat had always used a single, global RNG state that would be silently overwritten by each new thermalized bond. The new API no longer gives the false impression that each bond maintains its own RNG. The thermostat infrastructure is likely going to change in the near future anyway.

We should really consider rewriting the DPD thermostat to no longer get incremented by external components like constraints. This would allow us to have a single shared RNG counter from which all thermostats could extract their random numbers (the RNG salt decorrelates them). Then the DPD non-bonded potential and thermalized bonds would behave like any other interaction, since they would just fetch the RNG values from the system thermostat. This would also make the thermostat script interface a lot more maintainable. Right now it's quite convoluted, because the specificities of each thermostat prevent me from applying the SOLID principles. In particular, when turning a thermostat on, the arguments are dispatched to the thermostat container (kT) and to the actual thermostat (gamma, seed, etc.), which means that throwing a C++ exception requires rolling back the state of two script interface objects (both of which interact with the system class through distinct event code paths during that rollback) to gracefully restore the system into a valid state. I tried to refactor things further, but it proved to be a can of worm, and I could not find a solution that would both preserve the current API and play nice with the checkpointing mechanism.

In any case, I think this refactoring is a big improvement over the original Cython code, and from my side can be reviewed and merged as is. We can then consult with our users during the next software meeting how we want to proceed with further refactoring of the thermostats, and decide on which API changes would be acceptable. I would prefer not to delay the merge decision for this PR, because it simplifies the LB particle coupling code quite a bit and removes highly coupled global variables, and I would like to leverage that in my ongoing GPU LB work resp. ongoing globals removal work.

[jngrad]

There might actually be a bug in the LB particle coupling RNG. I have produced a MWE in #4848.